Supervised learning techniques for dairy cattle body weight prediction from 3D digital images

Area of Science:

• Agricultural Science
• Animal Science
• Computer Science

Background:

• Real-time monitoring of individual cows in livestock production is crucial for early anomaly detection and management.
• Body weight (BW) is a key indicator of productivity and health status in dairy cows.
• Automation and sensor-based systems offer advanced monitoring capabilities.

Purpose of the Study:

• To evaluate the performance of various supervised learning techniques for predicting dairy cow body weight (BW) using 3D image data.
• To compare prediction accuracies across different machine learning methods and datasets (single-breed vs. multi-breed).
• To assess the impact of data size and breed diversity on prediction performance.

Main Methods:

• Implemented and compared multiple supervised learning techniques (e.g., Catboost, AdaBoost, random forest) using 3D image contour data and BW measurements.
• Utilized a dataset of 83,011 records from 914 Danish Holstein and Jersey cows across 3 herds.
• Evaluated performance using Pearson's correlation coefficient (r), Root Mean Squared Error (RMSE), and Mean Absolute Percentage Error (MAPE).

Main Results:

• Achieved high prediction accuracies, with a mean correlation coefficient (r) up to 0.94 and low errors (MAPE as low as 4.0%, RMSE as low as 33.0 kg).
• Combined multi-breed datasets yielded superior predictive performance compared to single-breed analyses.
• Tree-based supervised learning techniques demonstrated the highest prediction accuracy across all metrics.

Conclusions:

• 3D image contour data combined with supervised learning, particularly tree-based methods, offers a promising approach for accurate dairy cow BW prediction.
• Increased data size, rather than multi-breed diversity alone, significantly enhances prediction performance.
• The developed methods show potential for application in commercial farm settings for improved livestock management.